Exploring the inherent fault tolerance of successive approximation algorithms under laser fault injection

International audienceThis work explores the fault tolerance of successive approximation algorithms, which are based on loop computations that approximate to a final result on each iteration. This type of approximate computing algorithms can present an inherent fault tolerance as they can manage small discrepancies in data values and converge to a final correct data after a certain number of interactions. A set of algorithms were implemented as embedded software in the ARM Cortex A9 processor of Xilinx Zynq-7000 series board. Experiments consist of exposing the finned processor to laser beams at a frequency of 10Hz. The laser beam travels through the data cache memory area provoking bit-flips. A computer connected to the device under test evaluates the execution outputs, checking for errors and classifying them into single and multiple silent data corruptions and hangs. Results show that by increasing the number of computational loops, it is possible to reduce the error rate significantly. They also show that most of the silent data corruption errors are not significant and can be accepted as correct by merely tolerating a result variation as little as 1%